It was only a few weeks since Amazon announced that it would offer AMD Epyc servers as an alternative to Amazon Web Services (AWS). Now, the company announces a new type of hardware platform for its newly created A1
instances – the Amazon Graviton processor.
The company writes:
Today, we are launching EC2 instances driven by arm-based AWS Graviton processors. Built around armored cores and extensive use of custom-built silicon, the A1 cases are optimized for performance and cost. They are great for extracting workloads, where you can share a group of smaller instances. It includes containerized micro services, web servers, development environments and caching fleets.
 Instances using script code can move their applications to A1 without rewriting, but if your code is running, you must rebuild it for an A1 instance.
Here’s what we know about Graviton so far. It is based on Cortex-A72, with a maximum clock speed of 2.3 GHz. AWS VP James Hamilton writes:
These new instances have up to 45% lower costs and will connect to 170 different types of instances supported by AWS, ranging from Intel-based z1d instances that provide a sustainable core rate of 4.0 GHz, a 12 TB memory instance, the F1 instance family with up to 8 field programmable gateways, P3 instances with the NVIDIA Tesla V100 GPU and the new M5a and R5a instances with AMD EPYC processors. No other cloud offer even comes close.
The new AWS-designed Arm-based A1 instances are available in 5 different instances from 1 core with 2 GiB memory up to 16 kernels with 32 GiB memory.
It is not clear exactly what kind of custom work Amazon did on the CPU. The register, the 16 vCPUs that make up each SoC, are arranged in clusters of four, with a 2MB L2 cache shared between each quad. Each single core has a 32KB L1 database and 48KB L1 instruction cache, which is a standard Cortex-A72 A vCPU map to a physical CPU core. Registe r reports that overall performance is quite varied in benchmarks. In some cases, 16 Graviton cores could not match 5 cores in a Xeon E5-2697v4, a Broadwell CPU.
It’s easy to forget, but once AMD makes a huge effort on ARM CPU, not x86. When AMD announced its K12 CPU back 2015, it chose to deliver “K” monics – a label previously reserved for x86 chips – to a future ARM kernel. K12, we were told, would share extensive resources and development strategy with Zen. There were plans for an ambitious common x86-ARM platform, called Project Skybridge. According to the registry, Amazon and AMD worked close to each other in 2015, until “AMD failed to meet all Amazon’s prestigious Amazon states.”
It is not clear which milestones, just AMD failed to meet – but we could choose from several. The company’s project Skybridge was suspended suddenly (we were wondering at the time that GF’s production problems could have been a part of the problem), and the Cortex-A57 CPU, Opteron A1100, was announced in 2014 but did not actually ship until 2016. If we were had to guess we would guess that the problems were with the A1100. Then AMD explained that part of the reason that the A1100 was held back, because the ARM server deployment infrastructure was not as solid as it needed to be and that more work needed to put the software in the stack. It may well have been true – ARM servers have taken years longer to come to market than originally expected – but Amazon was obviously not willing to wait. The company released AMD and purchased Annapurna Labs to carry out its design work.
This can also explain why AMD held K12. Lisa Su has focused on semi-custom parts, and the service as a dedicated supplier for Amazon ARM business could definitely have matched that goal. In this story, either Skybridge, Seattle or both trial balloons showed that AMD could bring ARM IP to the market before launching a new custom part based on some of the same architectural building blocks as its x86 CPU. When AMD lost the Amazon opportunity, it may have moved to its ARM design – and in turn, it may have been a part of why Jim Keller left the company when he did. All this is speculation, to be clear – it just fits the timeline quite well.
Regardless of how it turned out to be AMD, it’s the first time we’ve seen ARM servers used in cloud simulation and an important step forward. Amazon launches Alexa App in Windows Store, Amazon can split HQ2: One in Virginia, maybe in New York and Amazon tried to sell ICE’s Incorrect Face Detection Technology